sys/vfs/hammer2/hammer2_vfsops.c

   1 /*
   2  * Copyright (c) 2011-2015 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression)
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  *
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in
  16  *    the documentation and/or other materials provided with the
  17  *    distribution.
  18  * 3. Neither the name of The DragonFly Project nor the names of its
  19  *    contributors may be used to endorse or promote products derived
  20  *    from this software without specific, prior written permission.
  21  *
  22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  */
  35 #include <sys/param.h>
  36 #include <sys/systm.h>
  37 #include <sys/kernel.h>
  38 #include <sys/nlookup.h>
  39 #include <sys/vnode.h>
  40 #include <sys/mount.h>
  41 #include <sys/fcntl.h>
  42 #include <sys/buf.h>
  43 #include <sys/uuid.h>
  44 #include <sys/vfsops.h>
  45 #include <sys/sysctl.h>
  46 #include <sys/socket.h>
  47 #include <sys/objcache.h>
  48
  49 #include <sys/proc.h>
  50 #include <sys/namei.h>
  51 #include <sys/mountctl.h>
  52 #include <sys/dirent.h>
  53 #include <sys/uio.h>
  54
  55 #include <sys/mutex.h>
  56 #include <sys/mutex2.h>
  57
  58 #include "hammer2.h"
  59 #include "hammer2_disk.h"
  60 #include "hammer2_mount.h"
  61 #include "hammer2_lz4.h"
  62
  63 #include "zlib/hammer2_zlib.h"
  64
  65 #define REPORT_REFS_ERRORS 1    /* XXX remove me */
  66
  67 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
  68
  69 struct hammer2_sync_info {
  70         int error;
  71         int waitfor;
  72 };
  73
  74 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
  75 static struct hammer2_mntlist hammer2_mntlist;
  76
  77 struct hammer2_pfslist hammer2_pfslist;
  78 struct lock hammer2_mntlk;
  79
  80 int hammer2_debug;
  81 int hammer2_cluster_read = 4;           /* physical read-ahead */
  82 int hammer2_cluster_write = 0;          /* bdwrite() so later inval works */
  83 int hammer2_dedup_enable = 1;
  84 int hammer2_inval_enable = 0;
  85 int hammer2_flush_pipe = 100;
  86 int hammer2_synchronous_flush = 1;
  87 int hammer2_dio_count;
  88 long hammer2_chain_allocs;
  89 long hammer2_chain_frees;
  90 long hammer2_limit_dirty_chains;
  91 long hammer2_count_modified_chains;
  92 long hammer2_iod_invals;
  93 long hammer2_iod_file_read;
  94 long hammer2_iod_meta_read;
  95 long hammer2_iod_indr_read;
  96 long hammer2_iod_fmap_read;
  97 long hammer2_iod_volu_read;
  98 long hammer2_iod_file_write;
  99 long hammer2_iod_file_wembed;
 100 long hammer2_iod_file_wzero;
 101 long hammer2_iod_file_wdedup;
 102 long hammer2_iod_meta_write;
 103 long hammer2_iod_indr_write;
 104 long hammer2_iod_fmap_write;
 105 long hammer2_iod_volu_write;
 106
 107 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
 108 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
 109                 "Buffer used for compression.");
 110
 111 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
 112 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
 113                 "Buffer used for decompression.");
 114
 115 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
 116
 117 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
 118            &hammer2_debug, 0, "");
 119 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_read, CTLFLAG_RW,
 120            &hammer2_cluster_read, 0, "");
 121 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW,
 122            &hammer2_cluster_write, 0, "");
 123 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
 124            &hammer2_dedup_enable, 0, "");
 125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
 126            &hammer2_inval_enable, 0, "");
 127 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
 128            &hammer2_flush_pipe, 0, "");
 129 SYSCTL_INT(_vfs_hammer2, OID_AUTO, synchronous_flush, CTLFLAG_RW,
 130            &hammer2_synchronous_flush, 0, "");
 131 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
 132            &hammer2_chain_allocs, 0, "");
 133 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
 134            &hammer2_chain_frees, 0, "");
 135 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
 136            &hammer2_limit_dirty_chains, 0, "");
 137 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
 138            &hammer2_count_modified_chains, 0, "");
 139 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
 140            &hammer2_dio_count, 0, "");
 141
 142 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
 143            &hammer2_iod_invals, 0, "");
 144 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
 145            &hammer2_iod_file_read, 0, "");
 146 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
 147            &hammer2_iod_meta_read, 0, "");
 148 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
 149            &hammer2_iod_indr_read, 0, "");
 150 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
 151            &hammer2_iod_fmap_read, 0, "");
 152 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
 153            &hammer2_iod_volu_read, 0, "");
 154
 155 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
 156            &hammer2_iod_file_write, 0, "");
 157 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
 158            &hammer2_iod_file_wembed, 0, "");
 159 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
 160            &hammer2_iod_file_wzero, 0, "");
 161 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
 162            &hammer2_iod_file_wdedup, 0, "");
 163 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
 164            &hammer2_iod_meta_write, 0, "");
 165 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
 166            &hammer2_iod_indr_write, 0, "");
 167 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
 168            &hammer2_iod_fmap_write, 0, "");
 169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
 170            &hammer2_iod_volu_write, 0, "");
 171
 172 long hammer2_check_icrc32;
 173 long hammer2_check_xxhash64;
 174 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_icrc32, CTLFLAG_RW,
 175            &hammer2_check_icrc32, 0, "");
 176 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_xxhash64, CTLFLAG_RW,
 177            &hammer2_check_xxhash64, 0, "");
 178
 179 static int hammer2_vfs_init(struct vfsconf *conf);
 180 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
 181 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 182                                 struct ucred *cred);
 183 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
 184                                 struct vnode *, struct ucred *);
 185 static int hammer2_recovery(hammer2_dev_t *hmp);
 186 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
 187 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
 188 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
 189                                 struct ucred *cred);
 190 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
 191                                 struct ucred *cred);
 192 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
 193                                 struct fid *fhp, struct vnode **vpp);
 194 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
 195 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
 196                                 int *exflagsp, struct ucred **credanonp);
 197
 198 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
 199 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
 200
 201 static void hammer2_update_pmps(hammer2_dev_t *hmp);
 202
 203 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
 204 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
 205                                 hammer2_dev_t *hmp);
 206
 207 /*
 208  * HAMMER2 vfs operations.
 209  */
 210 static struct vfsops hammer2_vfsops = {
 211         .vfs_init       = hammer2_vfs_init,
 212         .vfs_uninit     = hammer2_vfs_uninit,
 213         .vfs_sync       = hammer2_vfs_sync,
 214         .vfs_mount      = hammer2_vfs_mount,
 215         .vfs_unmount    = hammer2_vfs_unmount,
 216         .vfs_root       = hammer2_vfs_root,
 217         .vfs_statfs     = hammer2_vfs_statfs,
 218         .vfs_statvfs    = hammer2_vfs_statvfs,
 219         .vfs_vget       = hammer2_vfs_vget,
 220         .vfs_vptofh     = hammer2_vfs_vptofh,
 221         .vfs_fhtovp     = hammer2_vfs_fhtovp,
 222         .vfs_checkexp   = hammer2_vfs_checkexp
 223 };
 224
 225 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
 226
 227 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
 228 MODULE_VERSION(hammer2, 1);
 229
 230 static
 231 int
 232 hammer2_vfs_init(struct vfsconf *conf)
 233 {
 234         static struct objcache_malloc_args margs_read;
 235         static struct objcache_malloc_args margs_write;
 236         static struct objcache_malloc_args margs_vop;
 237
 238         int error;
 239
 240         error = 0;
 241
 242         if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
 243                 error = EINVAL;
 244         if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
 245                 error = EINVAL;
 246         if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
 247                 error = EINVAL;
 248
 249         if (error)
 250                 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
 251
 252         margs_read.objsize = 65536;
 253         margs_read.mtype = M_HAMMER2_DEBUFFER;
 254
 255         margs_write.objsize = 32768;
 256         margs_write.mtype = M_HAMMER2_CBUFFER;
 257
 258         margs_vop.objsize = sizeof(hammer2_xop_t);
 259         margs_vop.mtype = M_HAMMER2;
 260
 261         /*
 262          * Note thaht for the XOPS cache we want backing store allocations
 263          * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
 264          * confusion), so use the backing store function that does it.  This
 265          * means that initial XOPS objects are zerod but REUSED objects are
 266          * not.  So we are responsible for cleaning the object up sufficiently
 267          * for our needs before objcache_put()ing it back (typically just the
 268          * FIFO indices).
 269          */
 270         cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
 271                                 0, 1, NULL, NULL, NULL,
 272                                 objcache_malloc_alloc,
 273                                 objcache_malloc_free,
 274                                 &margs_read);
 275         cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
 276                                 0, 1, NULL, NULL, NULL,
 277                                 objcache_malloc_alloc,
 278                                 objcache_malloc_free,
 279                                 &margs_write);
 280         cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
 281                                 0, 1, NULL, NULL, NULL,
 282                                 objcache_malloc_alloc_zero,
 283                                 objcache_malloc_free,
 284                                 &margs_vop);
 285
 286
 287         lockinit(&hammer2_mntlk, "mntlk", 0, 0);
 288         TAILQ_INIT(&hammer2_mntlist);
 289         TAILQ_INIT(&hammer2_pfslist);
 290
 291         hammer2_limit_dirty_chains = maxvnodes / 10;
 292         if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
 293                 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
 294
 295         return (error);
 296 }
 297
 298 static
 299 int
 300 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
 301 {
 302         objcache_destroy(cache_buffer_read);
 303         objcache_destroy(cache_buffer_write);
 304         objcache_destroy(cache_xops);
 305         return 0;
 306 }
 307
 308 /*
 309  * Core PFS allocator.  Used to allocate or reference the pmp structure
 310  * for PFS cluster mounts and the spmp structure for media (hmp) structures.
 311  * The pmp can be passed in or loaded by this function using the chain and
 312  * inode data.
 313  *
 314  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
 315  * transactions.  Note that synchronization does not use this field.
 316  * (typically frontend operations and synchronization cannot run on the
 317  * same PFS node at the same time).
 318  *
 319  * XXX check locking
 320  */
 321 hammer2_pfs_t *
 322 hammer2_pfsalloc(hammer2_chain_t *chain,
 323                  const hammer2_inode_data_t *ripdata,
 324                  hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
 325 {
 326         hammer2_pfs_t *pmp;
 327         hammer2_inode_t *iroot;
 328         int count;
 329         int i;
 330         int j;
 331
 332         pmp = NULL;
 333
 334         /*
 335          * Locate or create the PFS based on the cluster id.  If ripdata
 336          * is NULL this is a spmp which is unique and is always allocated.
 337          *
 338          * If the device is mounted in local mode all PFSs are considered
 339          * independent and not part of any cluster (for debugging only).
 340          */
 341         if (ripdata) {
 342                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 343                         if (force_local != pmp->force_local)
 344                                 continue;
 345                         if (force_local == NULL &&
 346                             bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
 347                                  sizeof(pmp->pfs_clid)) == 0) {
 348                                         break;
 349                         } else if (force_local && pmp->pfs_names[0] &&
 350                             strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
 351                                         break;
 352                         }
 353                 }
 354         }
 355
 356         if (pmp == NULL) {
 357                 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
 358                 pmp->force_local = force_local;
 359                 hammer2_trans_manage_init(pmp);
 360                 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
 361                 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
 362                 lockinit(&pmp->lock, "pfslk", 0, 0);
 363                 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
 364                 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
 365                 spin_init(&pmp->xop_spin, "h2xop");
 366                 spin_init(&pmp->lru_spin, "h2lru");
 367                 RB_INIT(&pmp->inum_tree);
 368                 TAILQ_INIT(&pmp->sideq);
 369                 TAILQ_INIT(&pmp->lru_list);
 370                 spin_init(&pmp->list_spin, "hm2pfsalloc_list");
 371
 372                 /*
 373                  * Distribute backend operations to threads
 374                  */
 375                 for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
 376                         hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
 377
 378                 /*
 379                  * Save the last media transaction id for the flusher.  Set
 380                  * initial
 381                  */
 382                 if (ripdata)
 383                         pmp->pfs_clid = ripdata->meta.pfs_clid;
 384                 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
 385
 386                 /*
 387                  * The synchronization thread may start too early, make
 388                  * sure it stays frozen until we are ready to let it go.
 389                  * XXX
 390                  */
 391                 /*
 392                 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
 393                                          HAMMER2_THREAD_REMASTER;
 394                 */
 395         }
 396
 397         /*
 398          * Create the PFS's root inode and any missing XOP helper threads.
 399          */
 400         if ((iroot = pmp->iroot) == NULL) {
 401                 iroot = hammer2_inode_get(pmp, NULL, NULL, -1);
 402                 if (ripdata)
 403                         iroot->meta = ripdata->meta;
 404                 pmp->iroot = iroot;
 405                 hammer2_inode_ref(iroot);
 406                 hammer2_inode_unlock(iroot);
 407         }
 408
 409         /*
 410          * Stop here if no chain is passed in.
 411          */
 412         if (chain == NULL)
 413                 goto done;
 414
 415         /*
 416          * When a chain is passed in we must add it to the PFS's root
 417          * inode, update pmp->pfs_types[], and update the syncronization
 418          * threads.
 419          *
 420          * When forcing local mode, mark the PFS as a MASTER regardless.
 421          *
 422          * At the moment empty spots can develop due to removals or failures.
 423          * Ultimately we want to re-fill these spots but doing so might
 424          * confused running code. XXX
 425          */
 426         hammer2_inode_ref(iroot);
 427         hammer2_mtx_ex(&iroot->lock);
 428         j = iroot->cluster.nchains;
 429
 430         kprintf("add PFS to pmp %p[%d]\n", pmp, j);
 431
 432         if (j == HAMMER2_MAXCLUSTER) {
 433                 kprintf("hammer2_mount: cluster full!\n");
 434                 /* XXX fatal error? */
 435         } else {
 436                 KKASSERT(chain->pmp == NULL);
 437                 chain->pmp = pmp;
 438                 hammer2_chain_ref(chain);
 439                 iroot->cluster.array[j].chain = chain;
 440                 if (force_local)
 441                         pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
 442                 else
 443                         pmp->pfs_types[j] = ripdata->meta.pfs_type;
 444                 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
 445                 pmp->pfs_hmps[j] = chain->hmp;
 446
 447                 /*
 448                  * If the PFS is already mounted we must account
 449                  * for the mount_count here.
 450                  */
 451                 if (pmp->mp)
 452                         ++chain->hmp->mount_count;
 453
 454                 /*
 455                  * May have to fixup dirty chain tracking.  Previous
 456                  * pmp was NULL so nothing to undo.
 457                  */
 458                 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
 459                         hammer2_pfs_memory_inc(pmp);
 460                 ++j;
 461         }
 462         iroot->cluster.nchains = j;
 463
 464         /*
 465          * Update nmasters from any PFS inode which is part of the cluster.
 466          * It is possible that this will result in a value which is too
 467          * high.  MASTER PFSs are authoritative for pfs_nmasters and will
 468          * override this value later on.
 469          *
 470          * (This informs us of masters that might not currently be
 471          *  discoverable by this mount).
 472          */
 473         if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
 474                 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
 475         }
 476
 477         /*
 478          * Count visible masters.  Masters are usually added with
 479          * ripdata->meta.pfs_nmasters set to 1.  This detects when there
 480          * are more (XXX and must update the master inodes).
 481          */
 482         count = 0;
 483         for (i = 0; i < iroot->cluster.nchains; ++i) {
 484                 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
 485                         ++count;
 486         }
 487         if (pmp->pfs_nmasters < count)
 488                 pmp->pfs_nmasters = count;
 489
 490         /*
 491          * Create missing synchronization and support threads.
 492          *
 493          * Single-node masters (including snapshots) have nothing to
 494          * synchronize and do not require this thread.
 495          *
 496          * Multi-node masters or any number of soft masters, slaves, copy,
 497          * or other PFS types need the thread.
 498          *
 499          * Each thread is responsible for its particular cluster index.
 500          * We use independent threads so stalls or mismatches related to
 501          * any given target do not affect other targets.
 502          */
 503         for (i = 0; i < iroot->cluster.nchains; ++i) {
 504                 /*
 505                  * Single-node masters (including snapshots) have nothing
 506                  * to synchronize and will make direct xops support calls,
 507                  * thus they do not require this thread.
 508                  *
 509                  * Note that there can be thousands of snapshots.  We do not
 510                  * want to create thousands of threads.
 511                  */
 512                 if (pmp->pfs_nmasters <= 1 &&
 513                     pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
 514                         continue;
 515                 }
 516
 517                 /*
 518                  * Sync support thread
 519                  */
 520                 if (pmp->sync_thrs[i].td == NULL) {
 521                         hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL,
 522                                            "h2nod", i, -1,
 523                                            hammer2_primary_sync_thread);
 524                 }
 525         }
 526
 527         /*
 528          * Create missing Xop threads
 529          *
 530          * NOTE: We create helper threads for all mounted PFSs or any
 531          *       PFSs with 2+ nodes (so the sync thread can update them,
 532          *       even if not mounted).
 533          */
 534         if (pmp->mp || iroot->cluster.nchains >= 2)
 535                 hammer2_xop_helper_create(pmp);
 536
 537         hammer2_mtx_unlock(&iroot->lock);
 538         hammer2_inode_drop(iroot);
 539 done:
 540         return pmp;
 541 }
 542
 543 /*
 544  * Deallocate an element of a probed PFS.  If destroying and this is a
 545  * MASTER, adjust nmasters.
 546  *
 547  * This function does not physically destroy the PFS element in its device
 548  * under the super-root  (see hammer2_ioctl_pfs_delete()).
 549  */
 550 void
 551 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
 552 {
 553         hammer2_inode_t *iroot;
 554         hammer2_chain_t *chain;
 555         int j;
 556
 557         /*
 558          * Cleanup our reference on iroot.  iroot is (should) not be needed
 559          * by the flush code.
 560          */
 561         iroot = pmp->iroot;
 562         if (iroot) {
 563                 /*
 564                  * Stop synchronizing
 565                  *
 566                  * XXX flush after acquiring the iroot lock.
 567                  * XXX clean out the cluster index from all inode structures.
 568                  */
 569                 hammer2_thr_delete(&pmp->sync_thrs[clindex]);
 570
 571                 /*
 572                  * Remove the cluster index from the group.  If destroying
 573                  * the PFS and this is a master, adjust pfs_nmasters.
 574                  */
 575                 hammer2_mtx_ex(&iroot->lock);
 576                 chain = iroot->cluster.array[clindex].chain;
 577                 iroot->cluster.array[clindex].chain = NULL;
 578
 579                 switch(pmp->pfs_types[clindex]) {
 580                 case HAMMER2_PFSTYPE_MASTER:
 581                         if (destroying && pmp->pfs_nmasters > 0)
 582                                 --pmp->pfs_nmasters;
 583                         /* XXX adjust ripdata->meta.pfs_nmasters */
 584                         break;
 585                 default:
 586                         break;
 587                 }
 588                 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
 589
 590                 hammer2_mtx_unlock(&iroot->lock);
 591
 592                 /*
 593                  * Release the chain.
 594                  */
 595                 if (chain) {
 596                         atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 597                         hammer2_chain_drop(chain);
 598                 }
 599
 600                 /*
 601                  * Terminate all XOP threads for the cluster index.
 602                  */
 603                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 604                         hammer2_thr_delete(&pmp->xop_groups[j].thrs[clindex]);
 605         }
 606 }
 607
 608 /*
 609  * Destroy a PFS, typically only occurs after the last mount on a device
 610  * has gone away.
 611  */
 612 static void
 613 hammer2_pfsfree(hammer2_pfs_t *pmp)
 614 {
 615         hammer2_inode_t *iroot;
 616         hammer2_chain_t *chain;
 617         int i;
 618         int j;
 619
 620         /*
 621          * Cleanup our reference on iroot.  iroot is (should) not be needed
 622          * by the flush code.
 623          */
 624         TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
 625
 626         iroot = pmp->iroot;
 627         if (iroot) {
 628                 for (i = 0; i < iroot->cluster.nchains; ++i) {
 629                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 630                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 631                                 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
 632                 }
 633 #if REPORT_REFS_ERRORS
 634                 if (pmp->iroot->refs != 1)
 635                         kprintf("PMP->IROOT %p REFS WRONG %d\n",
 636                                 pmp->iroot, pmp->iroot->refs);
 637 #else
 638                 KKASSERT(pmp->iroot->refs == 1);
 639 #endif
 640                 /* ref for pmp->iroot */
 641                 hammer2_inode_drop(pmp->iroot);
 642                 pmp->iroot = NULL;
 643         }
 644
 645         /*
 646          * Cleanup chains remaining on LRU list.
 647          */
 648         kprintf("pfsfree: %p lrucount=%d\n", pmp, pmp->lru_count);
 649         while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
 650                 hammer2_chain_ref(chain);
 651                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 652                 hammer2_chain_drop(chain);
 653         }
 654
 655         /*
 656          * Free remaining pmp resources
 657          */
 658         kmalloc_destroy(&pmp->mmsg);
 659         kmalloc_destroy(&pmp->minode);
 660
 661         kfree(pmp, M_HAMMER2);
 662 }
 663
 664 /*
 665  * Remove all references to hmp from the pfs list.  Any PFS which becomes
 666  * empty is terminated and freed.
 667  *
 668  * XXX inefficient.
 669  */
 670 static void
 671 hammer2_pfsfree_scan(hammer2_dev_t *hmp)
 672 {
 673         hammer2_pfs_t *pmp;
 674         hammer2_inode_t *iroot;
 675         hammer2_chain_t *rchain;
 676         int didfreeze;
 677         int i;
 678         int j;
 679
 680 again:
 681         TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 682                 if ((iroot = pmp->iroot) == NULL)
 683                         continue;
 684                 if (hmp->spmp == pmp) {
 685                         kprintf("unmount hmp %p remove spmp %p\n",
 686                                 hmp, pmp);
 687                         hmp->spmp = NULL;
 688                 }
 689
 690                 /*
 691                  * Determine if this PFS is affected.  If it is we must
 692                  * freeze all management threads and lock its iroot.
 693                  *
 694                  * Freezing a management thread forces it idle, operations
 695                  * in-progress will be aborted and it will have to start
 696                  * over again when unfrozen, or exit if told to exit.
 697                  */
 698                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 699                         if (pmp->pfs_hmps[i] == hmp)
 700                                 break;
 701                 }
 702                 if (i != HAMMER2_MAXCLUSTER) {
 703                         /*
 704                          * Make sure all synchronization threads are locked
 705                          * down.
 706                          */
 707                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 708                                 if (pmp->pfs_hmps[i] == NULL)
 709                                         continue;
 710                                 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 711                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 712                                         hammer2_thr_freeze_async(
 713                                                 &pmp->xop_groups[j].thrs[i]);
 714                                 }
 715                         }
 716                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 717                                 if (pmp->pfs_hmps[i] == NULL)
 718                                         continue;
 719                                 hammer2_thr_freeze(&pmp->sync_thrs[i]);
 720                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 721                                         hammer2_thr_freeze(
 722                                                 &pmp->xop_groups[j].thrs[i]);
 723                                 }
 724                         }
 725
 726                         /*
 727                          * Lock the inode and clean out matching chains.
 728                          * Note that we cannot use hammer2_inode_lock_*()
 729                          * here because that would attempt to validate the
 730                          * cluster that we are in the middle of ripping
 731                          * apart.
 732                          *
 733                          * WARNING! We are working directly on the inodes
 734                          *          embedded cluster.
 735                          */
 736                         hammer2_mtx_ex(&iroot->lock);
 737
 738                         /*
 739                          * Remove the chain from matching elements of the PFS.
 740                          */
 741                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 742                                 if (pmp->pfs_hmps[i] != hmp)
 743                                         continue;
 744                                 hammer2_thr_delete(&pmp->sync_thrs[i]);
 745                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 746                                         hammer2_thr_delete(
 747                                                 &pmp->xop_groups[j].thrs[i]);
 748                                 }
 749                                 rchain = iroot->cluster.array[i].chain;
 750                                 iroot->cluster.array[i].chain = NULL;
 751                                 pmp->pfs_types[i] = 0;
 752                                 if (pmp->pfs_names[i]) {
 753                                         kfree(pmp->pfs_names[i], M_HAMMER2);
 754                                         pmp->pfs_names[i] = NULL;
 755                                 }
 756                                 if (rchain) {
 757                                         hammer2_chain_drop(rchain);
 758                                         /* focus hint */
 759                                         if (iroot->cluster.focus == rchain)
 760                                                 iroot->cluster.focus = NULL;
 761                                 }
 762                                 pmp->pfs_hmps[i] = NULL;
 763                         }
 764                         hammer2_mtx_unlock(&iroot->lock);
 765                         didfreeze = 1;  /* remaster, unfreeze down below */
 766                 } else {
 767                         didfreeze = 0;
 768                 }
 769
 770                 /*
 771                  * Cleanup trailing chains.  Gaps may remain.
 772                  */
 773                 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
 774                         if (pmp->pfs_hmps[i])
 775                                 break;
 776                 }
 777                 iroot->cluster.nchains = i + 1;
 778
 779                 /*
 780                  * If the PMP has no elements remaining we can destroy it.
 781                  * (this will transition management threads from frozen->exit).
 782                  */
 783                 if (iroot->cluster.nchains == 0) {
 784                         kprintf("unmount hmp %p last ref to PMP=%p\n",
 785                                 hmp, pmp);
 786                         hammer2_pfsfree(pmp);
 787                         goto again;
 788                 }
 789
 790                 /*
 791                  * If elements still remain we need to set the REMASTER
 792                  * flag and unfreeze it.
 793                  */
 794                 if (didfreeze) {
 795                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 796                                 if (pmp->pfs_hmps[i] == NULL)
 797                                         continue;
 798                                 hammer2_thr_remaster(&pmp->sync_thrs[i]);
 799                                 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 800                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 801                                         hammer2_thr_remaster(
 802                                                 &pmp->xop_groups[j].thrs[i]);
 803                                         hammer2_thr_unfreeze(
 804                                                 &pmp->xop_groups[j].thrs[i]);
 805                                 }
 806                         }
 807                 }
 808         }
 809 }
 810
 811 /*
 812  * Mount or remount HAMMER2 fileystem from physical media
 813  *
 814  *      mountroot
 815  *              mp              mount point structure
 816  *              path            NULL
 817  *              data            <unused>
 818  *              cred            <unused>
 819  *
 820  *      mount
 821  *              mp              mount point structure
 822  *              path            path to mount point
 823  *              data            pointer to argument structure in user space
 824  *                      volume  volume path (device@LABEL form)
 825  *                      hflags  user mount flags
 826  *              cred            user credentials
 827  *
 828  * RETURNS:     0       Success
 829  *              !0      error number
 830  */
 831 static
 832 int
 833 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 834                   struct ucred *cred)
 835 {
 836         struct hammer2_mount_info info;
 837         hammer2_pfs_t *pmp;
 838         hammer2_pfs_t *spmp;
 839         hammer2_dev_t *hmp;
 840         hammer2_dev_t *force_local;
 841         hammer2_key_t key_next;
 842         hammer2_key_t key_dummy;
 843         hammer2_key_t lhc;
 844         struct vnode *devvp;
 845         struct nlookupdata nd;
 846         hammer2_chain_t *parent;
 847         hammer2_chain_t *chain;
 848         hammer2_cluster_t *cluster;
 849         const hammer2_inode_data_t *ripdata;
 850         hammer2_blockref_t bref;
 851         struct file *fp;
 852         char devstr[MNAMELEN];
 853         size_t size;
 854         size_t done;
 855         char *dev;
 856         char *label;
 857         int ronly = 1;
 858         int error;
 859         int cache_index;
 860         int i;
 861
 862         hmp = NULL;
 863         pmp = NULL;
 864         dev = NULL;
 865         label = NULL;
 866         devvp = NULL;
 867         cache_index = -1;
 868
 869         kprintf("hammer2_mount\n");
 870
 871         if (path == NULL) {
 872                 /*
 873                  * Root mount
 874                  */
 875                 bzero(&info, sizeof(info));
 876                 info.cluster_fd = -1;
 877                 ksnprintf(devstr, sizeof(devstr), "%s",
 878                           mp->mnt_stat.f_mntfromname);
 879                 kprintf("hammer2_mount: root '%s'\n", devstr);
 880         } else {
 881                 /*
 882                  * Non-root mount or updating a mount
 883                  */
 884                 error = copyin(data, &info, sizeof(info));
 885                 if (error)
 886                         return (error);
 887
 888                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 889                 if (error)
 890                         return (error);
 891         }
 892
 893         /*
 894          * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA
 895          * if no label specified, based on the partition id.  Error out if no
 896          * partition id.  This is strictly a convenience to match the
 897          * default label created by newfs_hammer2, our preference is
 898          * that a label always be specified.
 899          */
 900         dev = devstr;
 901         label = strchr(devstr, '@');
 902         if (label && ((label + 1) - dev) > done)
 903                 return (EINVAL);
 904         if (label && label == devstr)
 905                 return (EINVAL);
 906         if (label == NULL || label[1] == 0) {
 907                 char slice;
 908
 909                 if (label == NULL)
 910                         label = devstr + strlen(devstr);
 911                 slice = label[-1];
 912                 switch(slice) {
 913                 case 'a':
 914                         label = "BOOT";
 915                         break;
 916                 case 'd':
 917                         label = "ROOT";
 918                         break;
 919                 default:
 920                         label = "DATA";
 921                         break;
 922                 }
 923         } else {
 924                 *label = '\0';
 925                 label++;
 926         }
 927
 928         kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n",
 929                 dev, label, (mp->mnt_flag & MNT_RDONLY));
 930
 931         if (mp->mnt_flag & MNT_UPDATE) {
 932                 /*
 933                  * Update mount.  Note that pmp->iroot->cluster is
 934                  * an inode-embedded cluster and thus cannot be
 935                  * directly locked.
 936                  *
 937                  * XXX HAMMER2 needs to implement NFS export via
 938                  *     mountctl.
 939                  */
 940                 pmp = MPTOPMP(mp);
 941                 pmp->hflags = info.hflags;
 942                 cluster = &pmp->iroot->cluster;
 943                 for (i = 0; i < cluster->nchains; ++i) {
 944                         if (cluster->array[i].chain == NULL)
 945                                 continue;
 946                         hmp = cluster->array[i].chain->hmp;
 947                         devvp = hmp->devvp;
 948                         error = hammer2_remount(hmp, mp, path,
 949                                                 devvp, cred);
 950                         if (error)
 951                                 break;
 952                 }
 953
 954                 return error;
 955         }
 956
 957         /*
 958          * HMP device mount
 959          *
 960          * If a path is specified and dev is not an empty string, lookup the
 961          * name and verify that it referes to a block device.
 962          *
 963          * If a path is specified and dev is an empty string we fall through
 964          * and locate the label in the hmp search.
 965          */
 966         if (path && *dev != 0) {
 967                 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
 968                 if (error == 0)
 969                         error = nlookup(&nd);
 970                 if (error == 0)
 971                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
 972                 nlookup_done(&nd);
 973         } else if (path == NULL) {
 974                 /* root mount */
 975                 cdev_t cdev = kgetdiskbyname(dev);
 976                 error = bdevvp(cdev, &devvp);
 977                 if (error)
 978                         kprintf("hammer2: cannot find '%s'\n", dev);
 979         } else {
 980                 /*
 981                  * We will locate the hmp using the label in the hmp loop.
 982                  */
 983                 error = 0;
 984         }
 985
 986         if (error == 0 && devvp) {
 987                 if (vn_isdisk(devvp, &error))
 988                         error = vfs_mountedon(devvp);
 989         }
 990
 991         /*
 992          * Determine if the device has already been mounted.  After this
 993          * check hmp will be non-NULL if we are doing the second or more
 994          * hammer2 mounts from the same device.
 995          */
 996         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
 997         if (devvp) {
 998                 /*
 999                  * Match the device
1000                  */
1001                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1002                         if (hmp->devvp == devvp)
1003                                 break;
1004                 }
1005         } else if (error == 0) {
1006                 /*
1007                  * Match the label to a pmp already probed.
1008                  */
1009                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1010                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1011                                 if (pmp->pfs_names[i] &&
1012                                     strcmp(pmp->pfs_names[i], label) == 0) {
1013                                         hmp = pmp->pfs_hmps[i];
1014                                         break;
1015                                 }
1016                         }
1017                         if (hmp)
1018                                 break;
1019                 }
1020                 if (hmp == NULL)
1021                         error = ENOENT;
1022         }
1023
1024         /*
1025          * Open the device if this isn't a secondary mount and construct
1026          * the H2 device mount (hmp).
1027          */
1028         if (hmp == NULL) {
1029                 hammer2_chain_t *schain;
1030                 hammer2_xid_t xid;
1031
1032                 if (error == 0 && vcount(devvp) > 0)
1033                         error = EBUSY;
1034
1035                 /*
1036                  * Now open the device
1037                  */
1038                 if (error == 0) {
1039                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1040                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1041                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
1042                         if (error == 0) {
1043                                 error = VOP_OPEN(devvp,
1044                                              (ronly ? FREAD : FREAD | FWRITE),
1045                                              FSCRED, NULL);
1046                         }
1047                         vn_unlock(devvp);
1048                 }
1049                 if (error && devvp) {
1050                         vrele(devvp);
1051                         devvp = NULL;
1052                 }
1053                 if (error) {
1054                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1055                         return error;
1056                 }
1057                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1058                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1059                 hmp->ronly = ronly;
1060                 hmp->devvp = devvp;
1061                 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1062                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1063                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1064                 RB_INIT(&hmp->iotree);
1065                 spin_init(&hmp->io_spin, "hm2mount_io");
1066                 spin_init(&hmp->list_spin, "hm2mount_list");
1067                 TAILQ_INIT(&hmp->flushq);
1068
1069                 lockinit(&hmp->vollk, "h2vol", 0, 0);
1070                 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1071                 lockinit(&hmp->bflock, "h2bflk", 0, 0);
1072
1073                 /*
1074                  * vchain setup. vchain.data is embedded.
1075                  * vchain.refs is initialized and will never drop to 0.
1076                  *
1077                  * NOTE! voldata is not yet loaded.
1078                  */
1079                 hmp->vchain.hmp = hmp;
1080                 hmp->vchain.refs = 1;
1081                 hmp->vchain.data = (void *)&hmp->voldata;
1082                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1083                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1084                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1085
1086                 hammer2_chain_core_init(&hmp->vchain);
1087                 /* hmp->vchain.u.xxx is left NULL */
1088
1089                 /*
1090                  * fchain setup.  fchain.data is embedded.
1091                  * fchain.refs is initialized and will never drop to 0.
1092                  *
1093                  * The data is not used but needs to be initialized to
1094                  * pass assertion muster.  We use this chain primarily
1095                  * as a placeholder for the freemap's top-level RBTREE
1096                  * so it does not interfere with the volume's topology
1097                  * RBTREE.
1098                  */
1099                 hmp->fchain.hmp = hmp;
1100                 hmp->fchain.refs = 1;
1101                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1102                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1103                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1104                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1105                 hmp->fchain.bref.methods =
1106                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1107                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1108
1109                 hammer2_chain_core_init(&hmp->fchain);
1110                 /* hmp->fchain.u.xxx is left NULL */
1111
1112                 /*
1113                  * Install the volume header and initialize fields from
1114                  * voldata.
1115                  */
1116                 error = hammer2_install_volume_header(hmp);
1117                 if (error) {
1118                         hammer2_unmount_helper(mp, NULL, hmp);
1119                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1120                         hammer2_vfs_unmount(mp, MNT_FORCE);
1121                         return error;
1122                 }
1123
1124                 /*
1125                  * Really important to get these right or flush will get
1126                  * confused.
1127                  */
1128                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1129                 kprintf("alloc spmp %p tid %016jx\n",
1130                         hmp->spmp, hmp->voldata.mirror_tid);
1131                 spmp = hmp->spmp;
1132
1133                 /*
1134                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1135                  * is inherited from the volume header.
1136                  */
1137                 xid = 0;
1138                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1139                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1140                 hmp->vchain.pmp = spmp;
1141                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1142                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1143                 hmp->fchain.pmp = spmp;
1144
1145                 /*
1146                  * First locate the super-root inode, which is key 0
1147                  * relative to the volume header's blockset.
1148                  *
1149                  * Then locate the root inode by scanning the directory keyspace
1150                  * represented by the label.
1151                  */
1152                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1153                 schain = hammer2_chain_lookup(&parent, &key_dummy,
1154                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1155                                       &cache_index, 0);
1156                 hammer2_chain_lookup_done(parent);
1157                 if (schain == NULL) {
1158                         kprintf("hammer2_mount: invalid super-root\n");
1159                         hammer2_unmount_helper(mp, NULL, hmp);
1160                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1161                         hammer2_vfs_unmount(mp, MNT_FORCE);
1162                         return EINVAL;
1163                 }
1164                 if (schain->error) {
1165                         kprintf("hammer2_mount: error %s reading super-root\n",
1166                                 hammer2_error_str(schain->error));
1167                         hammer2_chain_unlock(schain);
1168                         hammer2_chain_drop(schain);
1169                         schain = NULL;
1170                         hammer2_unmount_helper(mp, NULL, hmp);
1171                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1172                         hammer2_vfs_unmount(mp, MNT_FORCE);
1173                         return EINVAL;
1174                 }
1175
1176                 /*
1177                  * The super-root always uses an inode_tid of 1 when
1178                  * creating PFSs.
1179                  */
1180                 spmp->inode_tid = 1;
1181                 spmp->modify_tid = schain->bref.modify_tid + 1;
1182
1183                 /*
1184                  * Sanity-check schain's pmp and finish initialization.
1185                  * Any chain belonging to the super-root topology should
1186                  * have a NULL pmp (not even set to spmp).
1187                  */
1188                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1189                 KKASSERT(schain->pmp == NULL);
1190                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1191
1192                 /*
1193                  * Replace the dummy spmp->iroot with a real one.  It's
1194                  * easier to just do a wholesale replacement than to try
1195                  * to update the chain and fixup the iroot fields.
1196                  *
1197                  * The returned inode is locked with the supplied cluster.
1198                  */
1199                 cluster = hammer2_cluster_from_chain(schain);
1200                 hammer2_inode_drop(spmp->iroot);
1201                 spmp->iroot = NULL;
1202                 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1);
1203                 spmp->spmp_hmp = hmp;
1204                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1205                 spmp->pfs_hmps[0] = hmp;
1206                 hammer2_inode_ref(spmp->iroot);
1207                 hammer2_inode_unlock(spmp->iroot);
1208                 hammer2_cluster_unlock(cluster);
1209                 hammer2_cluster_drop(cluster);
1210                 schain = NULL;
1211                 /* leave spmp->iroot with one ref */
1212
1213                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1214                         error = hammer2_recovery(hmp);
1215                         /* XXX do something with error */
1216                 }
1217                 hammer2_update_pmps(hmp);
1218                 hammer2_iocom_init(hmp);
1219                 hammer2_bulkfree_init(hmp);
1220
1221                 /*
1222                  * Ref the cluster management messaging descriptor.  The mount
1223                  * program deals with the other end of the communications pipe.
1224                  *
1225                  * Root mounts typically do not supply one.
1226                  */
1227                 if (info.cluster_fd >= 0) {
1228                         fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1229                         if (fp) {
1230                                 hammer2_cluster_reconnect(hmp, fp);
1231                         } else {
1232                                 kprintf("hammer2_mount: bad cluster_fd!\n");
1233                         }
1234                 }
1235         } else {
1236                 spmp = hmp->spmp;
1237                 if (info.hflags & HMNT2_DEVFLAGS) {
1238                         kprintf("hammer2: Warning: mount flags pertaining "
1239                                 "to the whole device may only be specified "
1240                                 "on the first mount of the device: %08x\n",
1241                                 info.hflags & HMNT2_DEVFLAGS);
1242                 }
1243         }
1244
1245         /*
1246          * Force local mount (disassociate all PFSs from their clusters).
1247          * Used primarily for debugging.
1248          */
1249         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1250
1251         /*
1252          * Lookup the mount point under the media-localized super-root.
1253          * Scanning hammer2_pfslist doesn't help us because it represents
1254          * PFS cluster ids which can aggregate several named PFSs together.
1255          *
1256          * cluster->pmp will incorrectly point to spmp and must be fixed
1257          * up later on.
1258          */
1259         hammer2_inode_lock(spmp->iroot, 0);
1260         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1261         lhc = hammer2_dirhash(label, strlen(label));
1262         chain = hammer2_chain_lookup(&parent, &key_next,
1263                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1264                                      &cache_index, 0);
1265         while (chain) {
1266                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1267                     strcmp(label, chain->data->ipdata.filename) == 0) {
1268                         break;
1269                 }
1270                 chain = hammer2_chain_next(&parent, chain, &key_next,
1271                                             key_next,
1272                                             lhc + HAMMER2_DIRHASH_LOMASK,
1273                                             &cache_index, 0);
1274         }
1275         if (parent) {
1276                 hammer2_chain_unlock(parent);
1277                 hammer2_chain_drop(parent);
1278         }
1279         hammer2_inode_unlock(spmp->iroot);
1280
1281         /*
1282          * PFS could not be found?
1283          */
1284         if (chain == NULL) {
1285                 kprintf("hammer2_mount: PFS label not found\n");
1286                 hammer2_unmount_helper(mp, NULL, hmp);
1287                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1288                 hammer2_vfs_unmount(mp, MNT_FORCE);
1289
1290                 return EINVAL;
1291         }
1292
1293         /*
1294          * Acquire the pmp structure (it should have already been allocated
1295          * via hammer2_update_pmps() so do not pass cluster in to add to
1296          * available chains).
1297          *
1298          * Check if the cluster has already been mounted.  A cluster can
1299          * only be mounted once, use null mounts to mount additional copies.
1300          */
1301         ripdata = &chain->data->ipdata;
1302         bref = chain->bref;
1303         pmp = hammer2_pfsalloc(NULL, ripdata,
1304                                bref.modify_tid, force_local);
1305         hammer2_chain_unlock(chain);
1306         hammer2_chain_drop(chain);
1307
1308         /*
1309          * Finish the mount
1310          */
1311         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1312
1313         if (pmp->mp) {
1314                 kprintf("hammer2_mount: PFS already mounted!\n");
1315                 hammer2_unmount_helper(mp, NULL, hmp);
1316                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1317                 hammer2_vfs_unmount(mp, MNT_FORCE);
1318
1319                 return EBUSY;
1320         }
1321
1322         pmp->hflags = info.hflags;
1323         mp->mnt_flag |= MNT_LOCAL;
1324         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1325         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1326
1327         /*
1328          * required mount structure initializations
1329          */
1330         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1331         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1332
1333         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1334         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1335
1336         /*
1337          * Optional fields
1338          */
1339         mp->mnt_iosize_max = MAXPHYS;
1340
1341         /*
1342          * Connect up mount pointers.
1343          */
1344         hammer2_mount_helper(mp, pmp);
1345
1346         lockmgr(&hammer2_mntlk, LK_RELEASE);
1347
1348         /*
1349          * Finish setup
1350          */
1351         vfs_getnewfsid(mp);
1352         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1353         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1354         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1355
1356         if (path) {
1357                 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1358                           MNAMELEN - 1, &size);
1359                 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1360         } /* else root mount, already in there */
1361
1362         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1363         if (path) {
1364                 copyinstr(path, mp->mnt_stat.f_mntonname,
1365                           sizeof(mp->mnt_stat.f_mntonname) - 1,
1366                           &size);
1367         } else {
1368                 /* root mount */
1369                 mp->mnt_stat.f_mntonname[0] = '/';
1370         }
1371
1372         /*
1373          * Initial statfs to prime mnt_stat.
1374          */
1375         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1376
1377         return 0;
1378 }
1379
1380 /*
1381  * Scan PFSs under the super-root and create hammer2_pfs structures.
1382  */
1383 static
1384 void
1385 hammer2_update_pmps(hammer2_dev_t *hmp)
1386 {
1387         const hammer2_inode_data_t *ripdata;
1388         hammer2_chain_t *parent;
1389         hammer2_chain_t *chain;
1390         hammer2_blockref_t bref;
1391         hammer2_dev_t *force_local;
1392         hammer2_pfs_t *spmp;
1393         hammer2_pfs_t *pmp;
1394         hammer2_key_t key_next;
1395         int cache_index = -1;
1396
1397         /*
1398          * Force local mount (disassociate all PFSs from their clusters).
1399          * Used primarily for debugging.
1400          */
1401         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1402
1403         /*
1404          * Lookup mount point under the media-localized super-root.
1405          *
1406          * cluster->pmp will incorrectly point to spmp and must be fixed
1407          * up later on.
1408          */
1409         spmp = hmp->spmp;
1410         hammer2_inode_lock(spmp->iroot, 0);
1411         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1412         chain = hammer2_chain_lookup(&parent, &key_next,
1413                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1414                                          &cache_index, 0);
1415         while (chain) {
1416                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1417                         continue;
1418                 ripdata = &chain->data->ipdata;
1419                 bref = chain->bref;
1420                 kprintf("ADD LOCAL PFS: %s\n", ripdata->filename);
1421
1422                 pmp = hammer2_pfsalloc(chain, ripdata,
1423                                        bref.modify_tid, force_local);
1424                 chain = hammer2_chain_next(&parent, chain, &key_next,
1425                                            key_next, HAMMER2_KEY_MAX,
1426                                            &cache_index, 0);
1427         }
1428         if (parent) {
1429                 hammer2_chain_unlock(parent);
1430                 hammer2_chain_drop(parent);
1431         }
1432         hammer2_inode_unlock(spmp->iroot);
1433 }
1434
1435 static
1436 int
1437 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1438                 struct vnode *devvp, struct ucred *cred)
1439 {
1440         int error;
1441
1442         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1443                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1444                 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1445                 vn_unlock(devvp);
1446                 error = hammer2_recovery(hmp);
1447                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1448                 if (error == 0) {
1449                         VOP_CLOSE(devvp, FREAD, NULL);
1450                         hmp->ronly = 0;
1451                 } else {
1452                         VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1453                 }
1454                 vn_unlock(devvp);
1455         } else {
1456                 error = 0;
1457         }
1458         return error;
1459 }
1460
1461 static
1462 int
1463 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1464 {
1465         hammer2_pfs_t *pmp;
1466         int flags;
1467         int error = 0;
1468
1469         pmp = MPTOPMP(mp);
1470
1471         if (pmp == NULL)
1472                 return(0);
1473
1474         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1475
1476         /*
1477          * If mount initialization proceeded far enough we must flush
1478          * its vnodes and sync the underlying mount points.  Three syncs
1479          * are required to fully flush the filesystem (freemap updates lag
1480          * by one flush, and one extra for safety).
1481          */
1482         if (mntflags & MNT_FORCE)
1483                 flags = FORCECLOSE;
1484         else
1485                 flags = 0;
1486         if (pmp->iroot) {
1487                 error = vflush(mp, 0, flags);
1488                 if (error)
1489                         goto failed;
1490                 hammer2_vfs_sync(mp, MNT_WAIT);
1491                 hammer2_vfs_sync(mp, MNT_WAIT);
1492                 hammer2_vfs_sync(mp, MNT_WAIT);
1493         }
1494
1495         /*
1496          * Cleanup the frontend support XOPS threads
1497          */
1498         hammer2_xop_helper_cleanup(pmp);
1499
1500         if (pmp->mp)
1501                 hammer2_unmount_helper(mp, pmp, NULL);
1502
1503         error = 0;
1504 failed:
1505         lockmgr(&hammer2_mntlk, LK_RELEASE);
1506
1507         return (error);
1508 }
1509
1510 /*
1511  * Mount helper, hook the system mount into our PFS.
1512  * The mount lock is held.
1513  *
1514  * We must bump the mount_count on related devices for any
1515  * mounted PFSs.
1516  */
1517 static
1518 void
1519 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1520 {
1521         hammer2_cluster_t *cluster;
1522         hammer2_chain_t *rchain;
1523         int i;
1524
1525         mp->mnt_data = (qaddr_t)pmp;
1526         pmp->mp = mp;
1527
1528         /*
1529          * After pmp->mp is set we have to adjust hmp->mount_count.
1530          */
1531         cluster = &pmp->iroot->cluster;
1532         for (i = 0; i < cluster->nchains; ++i) {
1533                 rchain = cluster->array[i].chain;
1534                 if (rchain == NULL)
1535                         continue;
1536                 ++rchain->hmp->mount_count;
1537                 kprintf("hammer2_mount hmp=%p ++mount_count=%d\n",
1538                         rchain->hmp, rchain->hmp->mount_count);
1539         }
1540
1541         /*
1542          * Create missing Xop threads
1543          */
1544         hammer2_xop_helper_create(pmp);
1545 }
1546
1547 /*
1548  * Mount helper, unhook the system mount from our PFS.
1549  * The mount lock is held.
1550  *
1551  * If hmp is supplied a mount responsible for being the first to open
1552  * the block device failed and the block device and all PFSs using the
1553  * block device must be cleaned up.
1554  *
1555  * If pmp is supplied multiple devices might be backing the PFS and each
1556  * must be disconnected.  This might not be the last PFS using some of the
1557  * underlying devices.  Also, we have to adjust our hmp->mount_count
1558  * accounting for the devices backing the pmp which is now undergoing an
1559  * unmount.
1560  */
1561 static
1562 void
1563 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1564 {
1565         hammer2_cluster_t *cluster;
1566         hammer2_chain_t *rchain;
1567         struct vnode *devvp;
1568         int dumpcnt;
1569         int ronly;
1570         int i;
1571
1572         /*
1573          * If no device supplied this is a high-level unmount and we have to
1574          * to disconnect the mount, adjust mount_count, and locate devices
1575          * that might now have no mounts.
1576          */
1577         if (pmp) {
1578                 KKASSERT(hmp == NULL);
1579                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1580                 pmp->mp = NULL;
1581                 mp->mnt_data = NULL;
1582
1583                 /*
1584                  * After pmp->mp is cleared we have to account for
1585                  * mount_count.
1586                  */
1587                 cluster = &pmp->iroot->cluster;
1588                 for (i = 0; i < cluster->nchains; ++i) {
1589                         rchain = cluster->array[i].chain;
1590                         if (rchain == NULL)
1591                                 continue;
1592                         --rchain->hmp->mount_count;
1593                         kprintf("hammer2_unmount hmp=%p --mount_count=%d\n",
1594                                 rchain->hmp, rchain->hmp->mount_count);
1595                         /* scrapping hmp now may invalidate the pmp */
1596                 }
1597 again:
1598                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1599                         if (hmp->mount_count == 0) {
1600                                 hammer2_unmount_helper(NULL, NULL, hmp);
1601                                 goto again;
1602                         }
1603                 }
1604                 return;
1605         }
1606
1607         /*
1608          * Try to terminate the block device.  We can't terminate it if
1609          * there are still PFSs referencing it.
1610          */
1611         kprintf("hammer2_unmount hmp=%p mount_count=%d\n",
1612                 hmp, hmp->mount_count);
1613         if (hmp->mount_count)
1614                 return;
1615
1616         hammer2_bulkfree_uninit(hmp);
1617         hammer2_pfsfree_scan(hmp);
1618         hammer2_dev_exlock(hmp);        /* XXX order */
1619
1620         /*
1621          * Cycle the volume data lock as a safety (probably not needed any
1622          * more).  To ensure everything is out we need to flush at least
1623          * three times.  (1) The running of the sideq can dirty the
1624          * filesystem, (2) A normal flush can dirty the freemap, and
1625          * (3) ensure that the freemap is fully synchronized.
1626          *
1627          * The next mount's recovery scan can clean everything up but we want
1628          * to leave the filesystem in a 100% clean state on a normal unmount.
1629          */
1630 #if 0
1631         hammer2_voldata_lock(hmp);
1632         hammer2_voldata_unlock(hmp);
1633 #endif
1634         hammer2_iocom_uninit(hmp);
1635
1636         if ((hmp->vchain.flags | hmp->fchain.flags) &
1637             HAMMER2_CHAIN_FLUSH_MASK) {
1638                 kprintf("hammer2_unmount: chains left over "
1639                         "after final sync\n");
1640                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1641                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1642
1643                 if (hammer2_debug & 0x0010)
1644                         Debugger("entered debugger");
1645         }
1646
1647         KKASSERT(hmp->spmp == NULL);
1648
1649         /*
1650          * Finish up with the device vnode
1651          */
1652         if ((devvp = hmp->devvp) != NULL) {
1653                 ronly = hmp->ronly;
1654                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1655                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1656                 hmp->devvp = NULL;
1657                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1658                 vn_unlock(devvp);
1659                 vrele(devvp);
1660                 devvp = NULL;
1661         }
1662
1663         /*
1664          * Clear vchain/fchain flags that might prevent final cleanup
1665          * of these chains.
1666          */
1667         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1668                 atomic_add_long(&hammer2_count_modified_chains, -1);
1669                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1670                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1671         }
1672         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1673                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1674         }
1675
1676         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1677                 atomic_add_long(&hammer2_count_modified_chains, -1);
1678                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1679                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1680         }
1681         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1682                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1683         }
1684
1685         /*
1686          * Final drop of embedded freemap root chain to
1687          * clean up fchain.core (fchain structure is not
1688          * flagged ALLOCATED so it is cleaned out and then
1689          * left to rot).
1690          */
1691         hammer2_chain_drop(&hmp->fchain);
1692
1693         /*
1694          * Final drop of embedded volume root chain to clean
1695          * up vchain.core (vchain structure is not flagged
1696          * ALLOCATED so it is cleaned out and then left to
1697          * rot).
1698          */
1699         dumpcnt = 50;
1700         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1701         dumpcnt = 50;
1702         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1703         hammer2_dev_unlock(hmp);
1704         hammer2_chain_drop(&hmp->vchain);
1705
1706         hammer2_io_cleanup(hmp, &hmp->iotree);
1707         if (hmp->iofree_count) {
1708                 kprintf("io_cleanup: %d I/O's left hanging\n",
1709                         hmp->iofree_count);
1710         }
1711
1712         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1713         kmalloc_destroy(&hmp->mchain);
1714         kfree(hmp, M_HAMMER2);
1715 }
1716
1717 int
1718 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1719                  ino_t ino, struct vnode **vpp)
1720 {
1721         hammer2_xop_lookup_t *xop;
1722         hammer2_pfs_t *pmp;
1723         hammer2_inode_t *ip;
1724         hammer2_tid_t inum;
1725         int error;
1726
1727         inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1728
1729         error = 0;
1730         pmp = MPTOPMP(mp);
1731
1732         /*
1733          * Easy if we already have it cached
1734          */
1735         ip = hammer2_inode_lookup(pmp, inum);
1736         if (ip) {
1737                 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1738                 *vpp = hammer2_igetv(ip, &error);
1739                 hammer2_inode_unlock(ip);
1740                 hammer2_inode_drop(ip);         /* from lookup */
1741
1742                 return error;
1743         }
1744
1745         /*
1746          * Otherwise we have to find the inode
1747          */
1748         xop = hammer2_xop_alloc(pmp->iroot, 0);
1749         xop->lhc = inum;
1750         hammer2_xop_start(&xop->head, hammer2_xop_lookup);
1751         error = hammer2_xop_collect(&xop->head, 0);
1752
1753         if (error == 0) {
1754                 if (hammer2_cluster_rdata(&xop->head.cluster) == NULL) {
1755                         kprintf("vget: no collect error but also no rdata\n");
1756                         kprintf("xop %p\n", xop);
1757                         while ((hammer2_debug & 0x80000) == 0) {
1758                                 tsleep(xop, PCATCH, "wait", hz * 10);
1759                         }
1760                         ip = NULL;
1761                 } else {
1762                         ip = hammer2_inode_get(pmp, NULL, &xop->head.cluster, -1);
1763                 }
1764         }
1765         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1766
1767         if (ip) {
1768                 *vpp = hammer2_igetv(ip, &error);
1769                 hammer2_inode_unlock(ip);
1770         } else {
1771                 *vpp = NULL;
1772                 error = ENOENT;
1773         }
1774         return (error);
1775 }
1776
1777 static
1778 int
1779 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1780 {
1781         hammer2_pfs_t *pmp;
1782         struct vnode *vp;
1783         int error;
1784
1785         pmp = MPTOPMP(mp);
1786         if (pmp->iroot == NULL) {
1787                 *vpp = NULL;
1788                 return EINVAL;
1789         }
1790
1791         error = 0;
1792         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1793
1794         while (pmp->inode_tid == 0) {
1795                 hammer2_xop_ipcluster_t *xop;
1796                 hammer2_inode_meta_t *meta;
1797
1798                 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1799                 hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1800                 error = hammer2_xop_collect(&xop->head, 0);
1801
1802                 if (error == 0) {
1803                         meta = &xop->head.cluster.focus->data->ipdata.meta;
1804                         pmp->iroot->meta = *meta;
1805                         pmp->inode_tid = meta->pfs_inum + 1;
1806                         if (pmp->inode_tid < HAMMER2_INODE_START)
1807                                 pmp->inode_tid = HAMMER2_INODE_START;
1808                         pmp->modify_tid =
1809                                 xop->head.cluster.focus->bref.modify_tid + 1;
1810                         kprintf("PFS: Starting inode %jd\n",
1811                                 (intmax_t)pmp->inode_tid);
1812                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1813                                 pmp->inode_tid, pmp->modify_tid);
1814                         wakeup(&pmp->iroot);
1815
1816                         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1817
1818                         /*
1819                          * Prime the mount info.
1820                          */
1821                         hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1822                         break;
1823                 }
1824
1825                 /*
1826                  * Loop, try again
1827                  */
1828                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1829                 hammer2_inode_unlock(pmp->iroot);
1830                 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1831                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1832                 if (error == EINTR)
1833                         break;
1834         }
1835
1836         if (error) {
1837                 hammer2_inode_unlock(pmp->iroot);
1838                 *vpp = NULL;
1839         } else {
1840                 vp = hammer2_igetv(pmp->iroot, &error);
1841                 hammer2_inode_unlock(pmp->iroot);
1842                 *vpp = vp;
1843         }
1844
1845         return (error);
1846 }
1847
1848 /*
1849  * Filesystem status
1850  *
1851  * XXX incorporate ipdata->meta.inode_quota and data_quota
1852  */
1853 static
1854 int
1855 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1856 {
1857         hammer2_pfs_t *pmp;
1858         hammer2_dev_t *hmp;
1859         hammer2_blockref_t bref;
1860         int i;
1861
1862         /*
1863          * NOTE: iroot might not have validated the cluster yet.
1864          */
1865         pmp = MPTOPMP(mp);
1866
1867         mp->mnt_stat.f_files = 0;
1868         mp->mnt_stat.f_ffree = 0;
1869         mp->mnt_stat.f_blocks = 0;
1870         mp->mnt_stat.f_bfree = 0;
1871         mp->mnt_stat.f_bavail = 0;
1872
1873         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1874                 hmp = pmp->pfs_hmps[i];
1875                 if (hmp == NULL)
1876                         continue;
1877                 if (pmp->iroot->cluster.array[i].chain)
1878                         bref = pmp->iroot->cluster.array[i].chain->bref;
1879                 else
1880                         bzero(&bref, sizeof(bref));
1881
1882                 mp->mnt_stat.f_files = bref.embed.stats.inode_count;
1883                 mp->mnt_stat.f_ffree = 0;
1884                 mp->mnt_stat.f_blocks = hmp->voldata.allocator_size /
1885                                         mp->mnt_vstat.f_bsize;
1886                 mp->mnt_stat.f_bfree = hmp->voldata.allocator_free /
1887                                         mp->mnt_vstat.f_bsize;
1888                 mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1889
1890                 *sbp = mp->mnt_stat;
1891         }
1892         return (0);
1893 }
1894
1895 static
1896 int
1897 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1898 {
1899         hammer2_pfs_t *pmp;
1900         hammer2_dev_t *hmp;
1901         hammer2_blockref_t bref;
1902         int i;
1903
1904         /*
1905          * NOTE: iroot might not have validated the cluster yet.
1906          */
1907         pmp = MPTOPMP(mp);
1908
1909         mp->mnt_vstat.f_bsize = 0;
1910         mp->mnt_vstat.f_files = 0;
1911         mp->mnt_vstat.f_ffree = 0;
1912         mp->mnt_vstat.f_blocks = 0;
1913         mp->mnt_vstat.f_bfree = 0;
1914         mp->mnt_vstat.f_bavail = 0;
1915
1916         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1917                 hmp = pmp->pfs_hmps[i];
1918                 if (hmp == NULL)
1919                         continue;
1920                 if (pmp->iroot->cluster.array[i].chain)
1921                         bref = pmp->iroot->cluster.array[i].chain->bref;
1922                 else
1923                         bzero(&bref, sizeof(bref));
1924
1925                 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1926                 mp->mnt_vstat.f_files = bref.embed.stats.inode_count;
1927                 mp->mnt_vstat.f_ffree = 0;
1928                 mp->mnt_vstat.f_blocks = hmp->voldata.allocator_size /
1929                                         mp->mnt_vstat.f_bsize;
1930                 mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
1931                                         mp->mnt_vstat.f_bsize;
1932                 mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1933
1934                 *sbp = mp->mnt_vstat;
1935         }
1936         return (0);
1937 }
1938
1939 /*
1940  * Mount-time recovery (RW mounts)
1941  *
1942  * Updates to the free block table are allowed to lag flushes by one
1943  * transaction.  In case of a crash, then on a fresh mount we must do an
1944  * incremental scan of the last committed transaction id and make sure that
1945  * all related blocks have been marked allocated.
1946  *
1947  * The super-root topology and each PFS has its own transaction id domain,
1948  * so we must track PFS boundary transitions.
1949  */
1950 struct hammer2_recovery_elm {
1951         TAILQ_ENTRY(hammer2_recovery_elm) entry;
1952         hammer2_chain_t *chain;
1953         hammer2_tid_t sync_tid;
1954 };
1955
1956 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
1957
1958 struct hammer2_recovery_info {
1959         struct hammer2_recovery_list list;
1960         hammer2_tid_t   mtid;
1961         int     depth;
1962 };
1963
1964 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
1965                         hammer2_chain_t *parent,
1966                         struct hammer2_recovery_info *info,
1967                         hammer2_tid_t sync_tid);
1968
1969 #define HAMMER2_RECOVERY_MAXDEPTH       10
1970
1971 static
1972 int
1973 hammer2_recovery(hammer2_dev_t *hmp)
1974 {
1975         struct hammer2_recovery_info info;
1976         struct hammer2_recovery_elm *elm;
1977         hammer2_chain_t *parent;
1978         hammer2_tid_t sync_tid;
1979         hammer2_tid_t mirror_tid;
1980         int error;
1981         int cumulative_error = 0;
1982
1983         hammer2_trans_init(hmp->spmp, 0);
1984
1985         sync_tid = hmp->voldata.freemap_tid;
1986         mirror_tid = hmp->voldata.mirror_tid;
1987
1988         kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
1989         if (sync_tid >= mirror_tid) {
1990                 kprintf(" no recovery needed\n");
1991         } else {
1992                 kprintf(" freemap recovery %016jx-%016jx\n",
1993                         sync_tid + 1, mirror_tid);
1994         }
1995
1996         TAILQ_INIT(&info.list);
1997         info.depth = 0;
1998         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1999         cumulative_error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2000         hammer2_chain_lookup_done(parent);
2001
2002         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2003                 TAILQ_REMOVE(&info.list, elm, entry);
2004                 parent = elm->chain;
2005                 sync_tid = elm->sync_tid;
2006                 kfree(elm, M_HAMMER2);
2007
2008                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2009                 error = hammer2_recovery_scan(hmp, parent, &info,
2010                                               hmp->voldata.freemap_tid);
2011                 hammer2_chain_unlock(parent);
2012                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
2013                 if (error)
2014                         cumulative_error = error;
2015         }
2016         hammer2_trans_done(hmp->spmp);
2017
2018         return cumulative_error;
2019 }
2020
2021 static
2022 int
2023 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2024                       struct hammer2_recovery_info *info,
2025                       hammer2_tid_t sync_tid)
2026 {
2027         const hammer2_inode_data_t *ripdata;
2028         hammer2_chain_t *chain;
2029         hammer2_blockref_t bref;
2030         int cache_index;
2031         int cumulative_error = 0;
2032         int error;
2033         int first;
2034
2035         /*
2036          * Adjust freemap to ensure that the block(s) are marked allocated.
2037          */
2038         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2039                 hammer2_freemap_adjust(hmp, &parent->bref,
2040                                        HAMMER2_FREEMAP_DORECOVER);
2041         }
2042
2043         /*
2044          * Check type for recursive scan
2045          */
2046         switch(parent->bref.type) {
2047         case HAMMER2_BREF_TYPE_VOLUME:
2048                 /* data already instantiated */
2049                 break;
2050         case HAMMER2_BREF_TYPE_INODE:
2051                 /*
2052                  * Must instantiate data for DIRECTDATA test and also
2053                  * for recursion.
2054                  */
2055                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2056                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
2057                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2058                         /* not applicable to recovery scan */
2059                         hammer2_chain_unlock(parent);
2060                         return 0;
2061                 }
2062                 hammer2_chain_unlock(parent);
2063                 break;
2064         case HAMMER2_BREF_TYPE_INDIRECT:
2065                 /*
2066                  * Must instantiate data for recursion
2067                  */
2068                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2069                 hammer2_chain_unlock(parent);
2070                 break;
2071         case HAMMER2_BREF_TYPE_DIRENT:
2072         case HAMMER2_BREF_TYPE_DATA:
2073         case HAMMER2_BREF_TYPE_FREEMAP:
2074         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2075         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2076                 /* not applicable to recovery scan */
2077                 return 0;
2078                 break;
2079         default:
2080                 return EDOM;
2081         }
2082
2083         /*
2084          * Defer operation if depth limit reached or if we are crossing a
2085          * PFS boundary.
2086          */
2087         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2088                 struct hammer2_recovery_elm *elm;
2089
2090                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2091                 elm->chain = parent;
2092                 elm->sync_tid = sync_tid;
2093                 hammer2_chain_ref(parent);
2094                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2095                 /* unlocked by caller */
2096
2097                 return(0);
2098         }
2099
2100
2101         /*
2102          * Recursive scan of the last flushed transaction only.  We are
2103          * doing this without pmp assignments so don't leave the chains
2104          * hanging around after we are done with them.
2105          */
2106         cache_index = 0;
2107         chain = NULL;
2108         first = 1;
2109
2110         while (hammer2_chain_scan(parent, &chain, &bref,
2111                                   &first, &cache_index,
2112                                   HAMMER2_LOOKUP_NODATA) != NULL) {
2113                 /*
2114                  * If this is a leaf
2115                  */
2116                 if (chain == NULL) {
2117                         if (bref.mirror_tid > sync_tid) {
2118                                 hammer2_freemap_adjust(hmp, &bref,
2119                                                      HAMMER2_FREEMAP_DORECOVER);
2120                         }
2121                         continue;
2122                 }
2123
2124                 /*
2125                  * This may or may not be a recursive node.
2126                  */
2127                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2128                 if (bref.mirror_tid > sync_tid) {
2129                         ++info->depth;
2130                         error = hammer2_recovery_scan(hmp, chain,
2131                                                       info, sync_tid);
2132                         --info->depth;
2133                         if (error)
2134                                 cumulative_error = error;
2135                 }
2136
2137                 /*
2138                  * Flush the recovery at the PFS boundary to stage it for
2139                  * the final flush of the super-root topology.
2140                  */
2141                 if ((bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2142                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2143                         hammer2_flush(chain, HAMMER2_FLUSH_TOP);
2144                 }
2145         }
2146
2147         return cumulative_error;
2148 }
2149
2150 /*
2151  * Sync a mount point; this is called on a per-mount basis from the
2152  * filesystem syncer process periodically and whenever a user issues
2153  * a sync.
2154  */
2155 int
2156 hammer2_vfs_sync(struct mount *mp, int waitfor)
2157 {
2158         hammer2_xop_flush_t *xop;
2159         struct hammer2_sync_info info;
2160         hammer2_inode_t *iroot;
2161         hammer2_pfs_t *pmp;
2162         int flags;
2163         int error;
2164
2165         pmp = MPTOPMP(mp);
2166         iroot = pmp->iroot;
2167         KKASSERT(iroot);
2168         KKASSERT(iroot->pmp == pmp);
2169
2170         /*
2171          * We can't acquire locks on existing vnodes while in a transaction
2172          * without risking a deadlock.  This assumes that vfsync() can be
2173          * called without the vnode locked (which it can in DragonFly).
2174          * Otherwise we'd have to implement a multi-pass or flag the lock
2175          * failures and retry.
2176          *
2177          * The reclamation code interlocks with the sync list's token
2178          * (by removing the vnode from the scan list) before unlocking
2179          * the inode, giving us time to ref the inode.
2180          */
2181         /*flags = VMSC_GETVP;*/
2182         flags = 0;
2183         if (waitfor & MNT_LAZY)
2184                 flags |= VMSC_ONEPASS;
2185
2186         /*
2187          * Preflush the vnodes using a normal transaction before interlocking
2188          * with a flush transaction.  We do this to try to run as much of
2189          * the compression as possible outside the flush transaction.
2190          *
2191          * For efficiency do an async pass before making sure with a
2192          * synchronous pass on all related buffer cache buffers.
2193          */
2194         hammer2_trans_init(pmp, 0);
2195         info.error = 0;
2196         info.waitfor = MNT_NOWAIT;
2197         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2198         info.waitfor = MNT_WAIT;
2199         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2200         hammer2_trans_done(pmp);
2201
2202         /*
2203          * Start our flush transaction.  This does not return until all
2204          * concurrent transactions have completed and will prevent any
2205          * new transactions from running concurrently, except for the
2206          * buffer cache transactions.
2207          *
2208          * NOTE!  It is still possible for the paging code to push pages
2209          *        out via a UIO_NOCOPY hammer2_vop_write() during the main
2210          *        flush.
2211          */
2212         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2213         hammer2_inode_run_sideq(pmp);
2214
2215         info.error = 0;
2216         info.waitfor = MNT_NOWAIT;
2217         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2218         info.waitfor = MNT_WAIT;
2219         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2220         hammer2_bioq_sync(pmp);
2221
2222         /*
2223          * Use the XOP interface to concurrently flush all nodes to
2224          * synchronize the PFSROOT subtopology to the media.  A standard
2225          * end-of-scan ENOENT error indicates cluster sufficiency.
2226          *
2227          * Note that this flush will not be visible on crash recovery until
2228          * we flush the super-root topology in the next loop.
2229          *
2230          * XXX For now wait for all flushes to complete.
2231          */
2232         if (iroot) {
2233                 xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING);
2234                 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2235                 error = hammer2_xop_collect(&xop->head,
2236                                             HAMMER2_XOP_COLLECT_WAITALL);
2237                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2238                 if (error == ENOENT)
2239                         error = 0;
2240         } else {
2241                 error = 0;
2242         }
2243         hammer2_trans_done(pmp);
2244
2245         return (error);
2246 }
2247
2248 /*
2249  * Sync passes.
2250  *
2251  * Note that we ignore the tranasction mtid we got above.  Instead,
2252  * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2253  * transactions.
2254  */
2255 static int
2256 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2257 {
2258         struct hammer2_sync_info *info = data;
2259         hammer2_inode_t *ip;
2260         int error;
2261
2262         /*
2263          * Degenerate cases.  Note that ip == NULL typically means the
2264          * syncer vnode itself and we don't want to vclrisdirty() in that
2265          * situation.
2266          */
2267         ip = VTOI(vp);
2268         if (ip == NULL) {
2269                 return(0);
2270         }
2271         if (vp->v_type == VNON || vp->v_type == VBAD) {
2272                 vclrisdirty(vp);
2273                 return(0);
2274         }
2275
2276         /*
2277          * VOP_FSYNC will start a new transaction so replicate some code
2278          * here to do it inline (see hammer2_vop_fsync()).
2279          *
2280          * WARNING: The vfsync interacts with the buffer cache and might
2281          *          block, we can't hold the inode lock at that time.
2282          *          However, we MUST ref ip before blocking to ensure that
2283          *          it isn't ripped out from under us (since we do not
2284          *          hold a lock on the vnode).
2285          */
2286         hammer2_inode_ref(ip);
2287         if ((ip->flags & HAMMER2_INODE_MODIFIED) ||
2288             !RB_EMPTY(&vp->v_rbdirty_tree)) {
2289                 vfsync(vp, info->waitfor, 1, NULL, NULL);
2290                 if (ip->flags & (HAMMER2_INODE_RESIZED |
2291                                  HAMMER2_INODE_MODIFIED)) {
2292                         hammer2_inode_lock(ip, 0);
2293                         hammer2_inode_chain_sync(ip);
2294                         hammer2_inode_unlock(ip);
2295                 }
2296         }
2297         if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2298             RB_EMPTY(&vp->v_rbdirty_tree)) {
2299                 vclrisdirty(vp);
2300         }
2301
2302         hammer2_inode_drop(ip);
2303 #if 1
2304         error = 0;
2305         if (error)
2306                 info->error = error;
2307 #endif
2308         return(0);
2309 }
2310
2311 static
2312 int
2313 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2314 {
2315         hammer2_inode_t *ip;
2316
2317         KKASSERT(MAXFIDSZ >= 16);
2318         ip = VTOI(vp);
2319         fhp->fid_len = offsetof(struct fid, fid_data[16]);
2320         fhp->fid_ext = 0;
2321         ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2322         ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2323
2324         return 0;
2325 }
2326
2327 static
2328 int
2329 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2330                struct fid *fhp, struct vnode **vpp)
2331 {
2332         hammer2_pfs_t *pmp;
2333         hammer2_tid_t inum;
2334         int error;
2335
2336         pmp = MPTOPMP(mp);
2337         inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2338         if (vpp) {
2339                 if (inum == 1)
2340                         error = hammer2_vfs_root(mp, vpp);
2341                 else
2342                         error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2343         } else {
2344                 error = 0;
2345         }
2346         if (error)
2347                 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2348         return error;
2349 }
2350
2351 static
2352 int
2353 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2354                  int *exflagsp, struct ucred **credanonp)
2355 {
2356         hammer2_pfs_t *pmp;
2357         struct netcred *np;
2358         int error;
2359
2360         pmp = MPTOPMP(mp);
2361         np = vfs_export_lookup(mp, &pmp->export, nam);
2362         if (np) {
2363                 *exflagsp = np->netc_exflags;
2364                 *credanonp = &np->netc_anon;
2365                 error = 0;
2366         } else {
2367                 error = EACCES;
2368         }
2369         return error;
2370 }
2371
2372 /*
2373  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2374  * header into the HMP
2375  *
2376  * XXX read four volhdrs and use the one with the highest TID whos CRC
2377  *     matches.
2378  *
2379  * XXX check iCRCs.
2380  *
2381  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2382  *     nonexistant locations.
2383  *
2384  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2385  */
2386 static
2387 int
2388 hammer2_install_volume_header(hammer2_dev_t *hmp)
2389 {
2390         hammer2_volume_data_t *vd;
2391         struct buf *bp;
2392         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2393         int error_reported;
2394         int error;
2395         int valid;
2396         int i;
2397
2398         error_reported = 0;
2399         error = 0;
2400         valid = 0;
2401         bp = NULL;
2402
2403         /*
2404          * There are up to 4 copies of the volume header (syncs iterate
2405          * between them so there is no single master).  We don't trust the
2406          * volu_size field so we don't know precisely how large the filesystem
2407          * is, so depend on the OS to return an error if we go beyond the
2408          * block device's EOF.
2409          */
2410         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2411                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2412                               HAMMER2_VOLUME_BYTES, &bp);
2413                 if (error) {
2414                         brelse(bp);
2415                         bp = NULL;
2416                         continue;
2417                 }
2418
2419                 vd = (struct hammer2_volume_data *) bp->b_data;
2420                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2421                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2422                         brelse(bp);
2423                         bp = NULL;
2424                         continue;
2425                 }
2426
2427                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2428                         /* XXX: Reversed-endianness filesystem */
2429                         kprintf("hammer2: reverse-endian filesystem detected");
2430                         brelse(bp);
2431                         bp = NULL;
2432                         continue;
2433                 }
2434
2435                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2436                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2437                                       HAMMER2_VOLUME_ICRC0_SIZE);
2438                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2439                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2440                                        HAMMER2_VOLUME_ICRC1_SIZE);
2441                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2442                         kprintf("hammer2 volume header crc "
2443                                 "mismatch copy #%d %08x/%08x\n",
2444                                 i, crc0, crc);
2445                         error_reported = 1;
2446                         brelse(bp);
2447                         bp = NULL;
2448                         continue;
2449                 }
2450                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2451                         valid = 1;
2452                         hmp->voldata = *vd;
2453                         hmp->volhdrno = i;
2454                 }
2455                 brelse(bp);
2456                 bp = NULL;
2457         }
2458         if (valid) {
2459                 hmp->volsync = hmp->voldata;
2460                 error = 0;
2461                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2462                         kprintf("hammer2: using volume header #%d\n",
2463                                 hmp->volhdrno);
2464                 }
2465         } else {
2466                 error = EINVAL;
2467                 kprintf("hammer2: no valid volume headers found!\n");
2468         }
2469         return (error);
2470 }
2471
2472 /*
2473  * This handles hysteresis on regular file flushes.  Because the BIOs are
2474  * routed to a thread it is possible for an excessive number to build up
2475  * and cause long front-end stalls long before the runningbuffspace limit
2476  * is hit, so we implement hammer2_flush_pipe to control the
2477  * hysteresis.
2478  *
2479  * This is a particular problem when compression is used.
2480  */
2481 void
2482 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2483 {
2484         atomic_add_int(&pmp->count_lwinprog, 1);
2485 }
2486
2487 void
2488 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2489 {
2490         int lwinprog;
2491
2492         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2493         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2494             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2495                 atomic_clear_int(&pmp->count_lwinprog,
2496                                  HAMMER2_LWINPROG_WAITING);
2497                 wakeup(&pmp->count_lwinprog);
2498         }
2499         if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2500             (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2501                 atomic_clear_int(&pmp->count_lwinprog,
2502                                  HAMMER2_LWINPROG_WAITING0);
2503                 wakeup(&pmp->count_lwinprog);
2504         }
2505 }
2506
2507 void
2508 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2509 {
2510         int lwinprog;
2511         int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2512                                     HAMMER2_LWINPROG_WAITING0;
2513
2514         for (;;) {
2515                 lwinprog = pmp->count_lwinprog;
2516                 cpu_ccfence();
2517                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2518                         break;
2519                 tsleep_interlock(&pmp->count_lwinprog, 0);
2520                 atomic_set_int(&pmp->count_lwinprog, lwflag);
2521                 lwinprog = pmp->count_lwinprog;
2522                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2523                         break;
2524                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2525         }
2526 }
2527
2528 /*
2529  * Manage excessive memory resource use for chain and related
2530  * structures.
2531  */
2532 void
2533 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2534 {
2535         uint32_t waiting;
2536         uint32_t count;
2537         uint32_t limit;
2538 #if 0
2539         static int zzticks;
2540 #endif
2541
2542         /*
2543          * Atomic check condition and wait.  Also do an early speedup of
2544          * the syncer to try to avoid hitting the wait.
2545          */
2546         for (;;) {
2547                 waiting = pmp->inmem_dirty_chains;
2548                 cpu_ccfence();
2549                 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2550
2551                 limit = pmp->mp->mnt_nvnodelistsize / 10;
2552                 if (limit < hammer2_limit_dirty_chains)
2553                         limit = hammer2_limit_dirty_chains;
2554                 if (limit < 1000)
2555                         limit = 1000;
2556
2557 #if 0
2558                 if ((int)(ticks - zzticks) > hz) {
2559                         zzticks = ticks;
2560                         kprintf("count %ld %ld\n", count, limit);
2561                 }
2562 #endif
2563
2564                 /*
2565                  * Block if there are too many dirty chains present, wait
2566                  * for the flush to clean some out.
2567                  */
2568                 if (count > limit) {
2569                         tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2570                         if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2571                                                waiting,
2572                                        waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2573                                 speedup_syncer(pmp->mp);
2574                                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2575                                        "chnmem", hz);
2576                         }
2577                         continue;       /* loop on success or fail */
2578                 }
2579
2580                 /*
2581                  * Try to start an early flush before we are forced to block.
2582                  */
2583                 if (count > limit * 7 / 10)
2584                         speedup_syncer(pmp->mp);
2585                 break;
2586         }
2587 }
2588
2589 void
2590 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2591 {
2592         if (pmp) {
2593                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2594         }
2595 }
2596
2597 void
2598 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2599 {
2600         uint32_t waiting;
2601
2602         if (pmp == NULL)
2603                 return;
2604
2605         for (;;) {
2606                 waiting = pmp->inmem_dirty_chains;
2607                 cpu_ccfence();
2608                 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2609                                        waiting,
2610                                        (waiting - 1) &
2611                                         ~HAMMER2_DIRTYCHAIN_WAITING)) {
2612                         break;
2613                 }
2614         }
2615
2616         if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2617                 wakeup(&pmp->inmem_dirty_chains);
2618 }
2619
2620 /*
2621  * Debugging
2622  */
2623 void
2624 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2625 {
2626         hammer2_chain_t *scan;
2627         hammer2_chain_t *parent;
2628
2629         --*countp;
2630         if (*countp == 0) {
2631                 kprintf("%*.*s...\n", tab, tab, "");
2632                 return;
2633         }
2634         if (*countp < 0)
2635                 return;
2636         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2637                 tab, tab, "", pfx,
2638                 chain, chain->bref.type,
2639                 chain->bref.key, chain->bref.keybits,
2640                 chain->bref.mirror_tid);
2641
2642         kprintf("%*.*s      [%08x] (%s) refs=%d",
2643                 tab, tab, "",
2644                 chain->flags,
2645                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2646                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2647                 chain->refs);
2648
2649         parent = chain->parent;
2650         if (parent)
2651                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2652                         tab, tab, "",
2653                         parent, parent->flags, parent->refs);
2654         if (RB_EMPTY(&chain->core.rbtree)) {
2655                 kprintf("\n");
2656         } else {
2657                 kprintf(" {\n");
2658                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2659                         hammer2_dump_chain(scan, tab + 4, countp, 'a');
2660                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2661                         kprintf("%*.*s}(%s)\n", tab, tab, "",
2662                                 chain->data->ipdata.filename);
2663                 else
2664                         kprintf("%*.*s}\n", tab, tab, "");
2665         }
2666 }